Finishing OpenMP
Controlling thread affinity
OpenMP thread binding:
- bind thread to place (core)
- once bound, thread does not migrate
OpenMP parallel proc_bind clause:
#pragma omp parallel proc_bind(master): whole team runs on same place as master#pragma omp parallell proc_bind(close): whole team runs on place close to parent (spread is the opposite)
places:
- basic computing resource, usually hardware thread
- describe hierarchic system architecture to OpenMP
- env variable
OMP_PLACES: “{0,1,2,3},{4,5,6,7},{8:4},{12:4}”
- sets 4 places with 4 execution units
- the actual effect is implementation-dependent
busy-wait vs suspension:
- problem: threads wait all the time, at synchronisation barriers and critical sections
- busy wait: threads actively poll for some memory address, advantage is threads can quickly proceed when possible, but they waste resources while waiting
- suspension: threads de-schedule to waiting queue, so they don’t occupy resources, but they have to wake up to be used
- use variable
OMP_WAIT_POLICY, ACTIVE for mostly busy wait, PASSIVE for mostly suspension (but not every compiler might know)
Atomic operations:
- atomic operations can be mapped to hardware features, but can also be mapped to critical sections
- no guarantee on operational behaviour
- can still be expensive due to data transfer
So, OpenMP:
- simpler to use that system-facing approaches
- responsible for most organisational aspects
- facilitates experimentation with alternatives
- false directives can lead to wrong or nondeterministic behaviour
- easy to write, hard to find
- insight into impact is reduced
POSIX Threads
POSIX threads are low-level, system-facing.
Library-based (pthread.h), don’t need a compiler that understands special directives.
Thread creation:
- only one thread at program startup
- all threads created dynamically
- threads may terminate before process
- process terminates when all threads terminate
- any thread can wait for termination of any other thread (“joining”)
- terminated threads are zombies until joined
- functions:
pthread_create, pthread_join
- joinable can be joined, detached cannot be joined
Two classes of threads:
- library threads: OS unaware, whole process blocked when waiting for OS service, implementation can be more efficient
OS/kernel threads: scheduled by OS together with other processes, pre-emptive multithreading, allows overlapping computation and waiting